October 25, 2002 PHPCon 2002 1 Making the Case for PHP at Yahoo! Michael J. Radwin mradwin@yahoo.com http://public.yahoo.com/~radwin/talks/

October 25, 2002 PHPCon 2002 2 Speaker Info • Michael J. Radwin – engineer for Yahoo! since 1998 – technical lead for the Apache web server – co-leading the PHP crusade at Y! • Contact Info: – mradwin@yahoo.com – http://www.radwin.org/michael/

October 25, 2002 PHPCon 2002 3 Outline • Motivation • History: from proprietary to Open Source • Choosing a new server-side scripting language – what the ideal system would look like – languages we didn’t choose – why we picked PHP • Scaling PHP • Lessons learned

October 25, 2002 PHPCon 2002 4 Motivation What’s so special about Yahoo!?

October 25, 2002 PHPCon 2002 5 World’s Biggest Site • World’s most trafficked Internet destination – Nielsen//NetRatings 8/2002 • Users – 201M unique users – 93M active registered users • Pageviews – more than 1.5 billion a day

October 25, 2002 PHPCon 2002 6 Huge Production Network • 4500+ servers • 16 co-locations – USA: Sunnyvale, Santa Clara, San Diego, Washington DC, Dallas – Intl: England, Central America, South America, Taiwan, Hong Kong, Singapore, China, Australia, India, Japan, Korea

October 25, 2002 PHPCon 2002 7 Complicated Software • Site – 74 properties • mail, shopping, sports, news, games, pets, etc. – 25 int’l sites – 13 languages • Code – 8.1M lines of C/C++ – 3.0M lines of Perl – 612 developers

October 25, 2002 PHPCon 2002 8 More about Y! Server Software It didn’t start out so complex…

October 25, 2002 PHPCon 2002 9 Y! Server Software: 1994-1995 • FreeBSD 2.1 (on Intel x86) • Filo server and Filo pages – 676 lines of C – optimized for speed – HTML + ads • CGIs for “dynamic” content – Search & Suggest A Site • advertisements client/server – yRPC homegrown RPC Early Years Static Content

October 25, 2002 PHPCon 2002 10 Y! Server Software: 1996-1998 • FreeBSD 2.1 and 2.2 • Apache 1.1 • Lots of home-grown software – free stuff wouldn’t scale, immature • yScript1 page Dynamic content – similar to Apache SSI – HTML + ads + personalization – content via include & DBM files • advertisements client/server • UDB (user data base) – NFS-mounted flat files Dynamic Content Personalization

October 25, 2002 PHPCon 2002 11 • FreeBSD 4.1 – a few Solaris boxes (Mail, Geo) • Apache 1.3.x • yScript2 pages – like yScript1, but more powerful – interactive forms – business logic in C++ • mod_python (Maps, YP) • UDB goes client/server – yRPC homegrown RPC Y! Server Software: 1999-2000 Boom Years Communications, Commerce, Communities

October 25, 2002 PHPCon 2002 12 Tradeoffs: App Logic in C++ • Advantages – fast execution speed – strongly typed, mature language • Disadvantages – edit, compile, link, debug cycle – not conducive to rapid prototyping – too easy to make mistakes with memory

October 25, 2002 PHPCon 2002 13 web server Example: my.yahoo.com browser user database server user prefs ad server ads web server news, weather, sports scores, stock quotes yScript load balancer yRPC yRPC feeds feeds

October 25, 2002 PHPCon 2002 14 Yahoo! in 2002 Moving towards Open Source

October 25, 2002 PHPCon 2002 15 Yahoo!’s Open Source Paradox • Open Source software runs our business – Perl – Apache – FreeBSD – GCC (+ GNU toolset) • Yet we seem to build a lot of our own stuff, too – RPC – server-side page languages – databases

October 25, 2002 PHPCon 2002 16 Are We Re-inventing the Wheel? • When Y! started in ’94 – free stuff did not scale – too immature – small community • How about today? – performance – integration – legacy & inertia – “Not Invented Here” syndrome

October 25, 2002 PHPCon 2002 17 Costs of Proprietary Languages • Maintenance – 3 different variants – C++ bugs • Training overhead – engineers – design folks • No integration – authoring tools, DBs • Limited functionality – yScript2 lacks subroutines! yScript

October 25, 2002 PHPCon 2002 18 Moving to Open Source • Open Source tech eventually matures – Y! replaced Filo server with Apache in 1996 – replacing some DBM and Oracle with MySQL • Server-side languages natural next step – features, performance, integration, community • Y! is a cheap company – economic recession 2001-2002 – can’t afford to waste engineering resources

October 25, 2002 PHPCon 2002 19 Choosing a Language How we ended up picking PHP

October 25, 2002 PHPCon 2002 20 Language Criteria 1. C/C++ extensions 2. loops, conditionals 3. complex data-types 4. pleasant syntax 5. runs on FreeBSD 6. high performance 7. robust, sand-boxed 8. interpreted (or dynamically compiled) 9. low training costs 10. i18n support 11. clean separation of presentation/content/app semantics 12. doesn’t require CS degree to use

October 25, 2002 PHPCon 2002 21 Why not Apache mod_include? • Pros – built into Apache, easy to learn/use • Limited language (no loops, subroutines) • Doesn’t interface with Y! code – Ads, User Database, etc. • Poor performance – parses file every time you hit page

October 25, 2002 PHPCon 2002 22 Why not ASP or Cold Fusion? • Pros – lots of 3rd-party integration – professional support • Cons – CF has ugly syntax – $$ for languages – $$ for Microsoft Windows

October 25, 2002 PHPCon 2002 23 Why not Perl? • Pros – FreeBSD support and performance is great – huge CPAN library – we already use it for offline processing • Cons – There’s More Than One Way To Do It – poor sandboxing, easy to screw up server – wasn’t designed as web scripting language

October 25, 2002 PHPCon 2002 24 Why not JSP, Servlets, or J2EE? • Pros – strongly typed – good performance (JIT), sandboxing – works w/lots of off-the-shelf software • But… you can’t really use Java w/o threads • Threads support on FreeBSD is not great

October 25, 2002 PHPCon 2002 25 Why not XSLT or ClearSilver? • Pro: separates HTML presentation from app logic • XSLT – complicated to set up and understand • ClearSilver – small developer community • Neither is “procedural” language – totally different models from PHP/ASP/JSP/yScript2 – difficult transition for Y! engineering

October 25, 2002 PHPCon 2002 26 So Why Did We Pick PHP? 1. Designed for server side web scripting 2. Large, Open Source developer community • integration, libraries • documentation & training 3. Debugging & profiling tools 4. Simple and clear syntax (fits Y! paradigm) 5. Performs well in our tests • efficient (with acceleration) • small enough memory footprint

October 25, 2002 PHPCon 2002 27 Benchmarking PHP “But is it as fast as yScript2?”

October 25, 2002 PHPCon 2002 28 Performance Tests • Languages – PHP 4.1.2 (w/Accel) – yScript2 (proprietary) – YSP (mod_perl) • Hardware – Pentium III 800Mhz – 512 Mb RAM – FreeBSD 4.3

October 25, 2002 PHPCon 2002 29 Performance Tests • 33K input script, 41K output • Included and evaluated 3 other files – header, navbar, footer • Echoed environment variables • Pseudo-personalization – “Hello, mradwin” • Called external C++ library for Ads/UDB – network delay to fetch data

October 25, 2002 PHPCon 2002 30 Performance: Requests Requests/sec 0 50 100 150 200 250 300 350 25 50 75 100 150 200 300 400 500 Concurrent requests req/s PHP YSP HF2k Network max yScript2

October 25, 2002 PHPCon 2002 31 Performance: Transfer Rate Transfer Rate 0 2000 4000 6000 8000 10000 12000 14000 25 50 75 100 150 200 300 400 500 Concurrent requests transfer rate (kb/s) PHP YSP HF2k yScript2

October 25, 2002 PHPCon 2002 32 Performance: Processing Time Processing time 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 25 50 75 100 150 200 300 400 500 Concurrent requests ms PHP YSP HF2k yScript2

October 25, 2002 PHPCon 2002 33 Performance: Memory Active Virtual Memory 0 200000 400000 600000 800000 1000000 25 50 75 100 150 200 300 400 500 Concurrent requests kbytes active PHP YSP HF2k yScript2

October 25, 2002 PHPCon 2002 34 Performance: Scaling PHP • Profile your code foreach ($_SERVER as $k => $v) if (substr($k, 0, 5) == “HTTP_”) $str .= substr($k, 5) . “: $v\n”; versus: if (strncmp($k, “HTTP_”, 5) == 0) • Implement C and C++ extensions – when you’re willing to trade flexibility for speed • Use an Accelerator

October 25, 2002 PHPCon 2002 35 Lessons Learned 4 months after we started using PHP

October 25, 2002 PHPCon 2002 36 Early Adopters • PHP for new properties – remember.yahoo.com for Sep 11 2002 • Internal tools – content mgmt, package repository, aclviewer • Most Y! properties integrating slowly – no plans to rewrite entire site – mix PHP, Apache DSOs, yScript1 & yScript2 pages

October 25, 2002 PHPCon 2002 37 Coding PHP Takes Discipline • Shallow learning curve – very easy to get some pages up quickly • But mixed app/presentation problematic – PHP code and HTML forever intertwined – coding conventions help • *.inc for function and class libraries • *.php for web pages (call functions, echo $vars) – use Smarty to enforce separation?

October 25, 2002 PHPCon 2002 38 PHP != Perl • The “implement twice” problem – much offline processing done in Perl – example: tax/shipping calculation for Shopping • PEAR != CPAN – installer doesn’t work in PHP 4.2.x – repository smaller, less mature than CPAN • Surprises for people used to coding Perl

October 25, 2002 PHPCon 2002 39 Giving Back to Open Source • We customize Open Source software we use – often improvements are not sent back – many are gross Y!-specific hacks • Improving our relationship with OS community – FreeBSD (Peter Wemm) – Apache (Sander van Zoest) – PHP (Rasmus Lerdorf) – MySQL (Jeremy Zawodny)

October 25, 2002 PHPCon 2002 40 Questions and Answers Slides online at: http://public.yahoo.com/~radwin/talks/

October 25, 2002 PHPCon 2002 41 Legal Mumbo-Jumbo • Text of this presentation is Copyright © 2002 Michael J. Radwin. • Clip art is Copyright © 2002 Microsoft Corporation. • Yahoo!, the Yahoo! logo, the “Jumpin’ Y Guy” logo, and other Yahoo! logos, product & service names are trademarks of Yahoo! Inc. • The Yahoo! Engineering logo is Copyright © 2000 John “JR” Conlin. • The PHP logo is Copyright © 2001, 2002 The PHP Group. • The Open Source, Apache Feather, Active Server Pages, Cold Fusion, “Powered By FreeBSD”, mod_perl, Apache::ASP, Mason, Java, W3C, Neotonic, and ionCube logos are Copyright © their respective owners.